positive pe/atom, multi-element system: SNAP

[materialsvirtuallab]

positive pe/atom, multi-element system: SNAP

[chc273]

@Rana-Phy thanks for the comments. I think the optimization finds the solutions with one positive value for bulk energy to be the most probable solution. It is from the data you chose rather than the code mandating such behavior.

An easy test would be to copy the bulk elemental data in the training a few times. This is essentially adding a large weight to those data points to make sure that the model predicts the correct elemental energy.

[chc273]

@Rana-Phy it is also possible to constrain the snap coeff. Basically, the snap coeff is just the coefficients fitted from linear regression. You can put the constraints that you like fairly easily to the linear fitting part, but that requires you to modify some part in the code.

[Rana-Phy]

@chc273 Thanks for your quick reply. I have a quick check on my data set.
Could it come from the mixing Ti and Ti_sv versions (VASP potential)?
I have used Ti PAW for large structures (100-200 atom cases) for faster calculation.
Other DFT calculations are done with Ti_sv.
If there is any way or tips for energy correction manually due to PAW version mixing kindly let us know.

[chc273]

@Rana-Phy ah, this could be a problem. The abs energy can be different using different pseudopotentials. Please keep them consistent.

[Rana-Phy]

@chc273 It seems like fitsnap code has a tool for shifting energy.
https://github.com/FitSNAP/FitSNAP/tree/master/tools